What problem does it solve? When serving fp8 a8w8 block-scale models on AMD MI300X GPUs, config-level Triton tuning of dense GEMM heads plateaus at roughly 1.04-1.06x, leaving the K-heavy down-projection GEMM as a major GPU-time bottleneck with no further config-level gains available. ## Core Features & Use Cases - Fused fp8 GEMM authoring: Guides authoring a FlyDSL fused a8w8 blockscale GEMM core that folds block-scale into the operand scale and runs one full-K low-precision MFMA, achieving ~2.43x isolated speedup on the down-proj shape. - Capture-safe binding: Describes the bare-core bind overlay (nocgraph) that runs inside sglang's existing captured CUDA graph, avoiding the nested-graph capture crash that blocked earlier attempts. - Memory-safe caching: Constrains weight caching to compact fp8/preshuffled weights keyed by weight.data_ptr(), preventing KV-cache starvation and e2e regressions. - Use Case: Optimizing Qwen3.5-27B-FP8 serving on gfx942, where applying this playbook raised throughput from 931.6 to 1559.9 tok/s (+67.4%) with parity passing. ## Quick Start Ask the agent to apply the FlyDSL fused fp8 blockscale GEMM playbook to the down-projection head of your fp8 model on MI300X with a capture-safe bare-core bind.
